Sentry Skills memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

The strongest supported boundary is editable guidance with separate evidence and task consumers. RT-RTE-skill-revision and RT-OBJ-skill-md support changes intended for later loading; static installation alone is not accumulated memory. Markdown rules are natural-language at model consumers, while changed invocation metadata is symbolic at external host consumers. Neither form licenses compliance or observed activation.

MEM-OBJ-working-set and MEM-OBJ-findings-log distinguish semi-raw examples from interpreted causes/deltas. EPI-RTE-example-retention is the retention part of the revision route, not an additional copy of that route. Its summarization supports claimed consolidation; promotion into standing rules supports claimed promotion. The revision account also supports claimed evolution and removal of redundant rules. No particular novel proposition establishes synthesis, despite the source-native synthesis route name.

Experience has learning authority at the guidance-update consumer; the output has instruction authority at the later task consumer. The prescribed model-admitted example-to-guidance route supports trace_learning "yes" at claimed basis. This does not establish automatic ingestion, exercised update, improved capacity or successful self-improvement. Actual example origins and their original trace contents remain unknown, so trace_source has no positive controlled classification. The manual maintainer gate and explicit holdout reassignment remain separate admission mechanisms; neither converts every model-chosen delta to manual admission.

MEM-OBJ-sources retains reasons for maintenance, and MEM-OBJ-spec supplies an interpreted conformity contract. Runtime consumers are not promised those reasons. MEM-RTE-sred-handoff instead requests a prior summary as evidence; MEM-RTE-reader-testing requests a saved draft for comprehension. Both are pull chains, including the manual reader transfer. They are not automatically instruction-learning routes merely because models generate documents.

Host relevance activation supplies a claimed push path through RT-BAP-skill-consumption. Generated project defaults supply an afforded push through MEM-RTE-project-guidance. Neither accepted account establishes the automatic selector's actual matching algorithm. A filename, project scope or requested summary link alone does not warrant an identifier push signal. Requested step-reference reads remain pull rather than being counted again as push delivery.

The reconciliation resolves the epistemic and workflow reply duplicates to MEM-RTE-thread-reply, and the epistemic reader duplicate to MEM-RTE-reader-testing. This profile uses those canonical routes. Reader supersession preserves the fuller saved-draft read-back account without upgrading isolation or fact checking. Reply supersessions do not establish future agent receipt: service publication and first-comment/self-comment selection remain distinct. The evidence-file container is not superseded; this profile uses its separately recorded example/findings parts without double-counting them.

RT-RTE-maintainer-eval records a configured evaluation harness but opaque retained results. Wiring of scenario generation does not upgrade report substrate, original trace contents, baseline initialization or later payload delivery. The file alternative for a document draft does not classify its host-artifact alternative. These uncertainties and unenumerated domain leaves prevent system-complete axis inventories. They remain explicitly included rather than being discarded to obtain known coverage. No accepted evidenced-absence record warrants a system-wide negative.

No frozen-source files were read for this classification. All warrants come from supplied accepted records and their annotations; reconciliation was used only to resolve identity. No new records, quotes or evidence were added.